diff --git a/common/jpeg/simd/jcqntsse.asm b/common/jpeg/simd/jcqntsse.asm
new file mode 100644
index 0000000..da7a3ff
--- /dev/null
+++ b/common/jpeg/simd/jcqntsse.asm
@@ -0,0 +1,208 @@
+;
+; jcqntsse.asm - sample data conversion and quantization (SSE & MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "simd/jsimdext.inc"
+%include "simd/jdct.inc"
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Load data into workspace, applying unsigned->signed conversion
+;
+; GLOBAL(void)
+; jsimd_convsamp_float_sse (JSAMPARRAY sample_data, JDIMENSION start_col,
+;                           FAST_FLOAT * workspace);
+;
+
+%define sample_data	ebp+8		; JSAMPARRAY sample_data
+%define start_col	ebp+12		; JDIMENSION start_col
+%define workspace	ebp+16		; FAST_FLOAT * workspace
+
+	align	16
+	global	EXTN(jsimd_convsamp_float_sse)
+
+EXTN(jsimd_convsamp_float_sse):
+	push	ebp
+	mov	ebp,esp
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	pcmpeqw  mm7,mm7
+	psllw    mm7,7
+	packsswb mm7,mm7		; mm7 = PB_CENTERJSAMPLE (0x808080..)
+
+	mov	esi, JSAMPARRAY [sample_data]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [start_col]
+	mov	edi, POINTER [workspace]	; (DCTELEM *)
+	mov	ecx, DCTSIZE/2
+	alignx	16,7
+.convloop:
+	mov	ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+	mov	edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW]	; (JSAMPLE *)
+
+	movq	mm0, MMWORD [ebx+eax*SIZEOF_JSAMPLE]
+	movq	mm1, MMWORD [edx+eax*SIZEOF_JSAMPLE]
+
+	psubb	mm0,mm7				; mm0=(01234567)
+	psubb	mm1,mm7				; mm1=(89ABCDEF)
+
+	punpcklbw mm2,mm0			; mm2=(*0*1*2*3)
+	punpckhbw mm0,mm0			; mm0=(*4*5*6*7)
+	punpcklbw mm3,mm1			; mm3=(*8*9*A*B)
+	punpckhbw mm1,mm1			; mm1=(*C*D*E*F)
+
+	punpcklwd mm4,mm2			; mm4=(***0***1)
+	punpckhwd mm2,mm2			; mm2=(***2***3)
+	punpcklwd mm5,mm0			; mm5=(***4***5)
+	punpckhwd mm0,mm0			; mm0=(***6***7)
+
+	psrad     mm4,(DWORD_BIT-BYTE_BIT)	; mm4=(01)
+	psrad     mm2,(DWORD_BIT-BYTE_BIT)	; mm2=(23)
+	cvtpi2ps  xmm0,mm4			; xmm0=(01**)
+	cvtpi2ps  xmm1,mm2			; xmm1=(23**)
+	psrad     mm5,(DWORD_BIT-BYTE_BIT)	; mm5=(45)
+	psrad     mm0,(DWORD_BIT-BYTE_BIT)	; mm0=(67)
+	cvtpi2ps  xmm2,mm5			; xmm2=(45**)
+	cvtpi2ps  xmm3,mm0			; xmm3=(67**)
+
+	punpcklwd mm6,mm3			; mm6=(***8***9)
+	punpckhwd mm3,mm3			; mm3=(***A***B)
+	punpcklwd mm4,mm1			; mm4=(***C***D)
+	punpckhwd mm1,mm1			; mm1=(***E***F)
+
+	psrad     mm6,(DWORD_BIT-BYTE_BIT)	; mm6=(89)
+	psrad     mm3,(DWORD_BIT-BYTE_BIT)	; mm3=(AB)
+	cvtpi2ps  xmm4,mm6			; xmm4=(89**)
+	cvtpi2ps  xmm5,mm3			; xmm5=(AB**)
+	psrad     mm4,(DWORD_BIT-BYTE_BIT)	; mm4=(CD)
+	psrad     mm1,(DWORD_BIT-BYTE_BIT)	; mm1=(EF)
+	cvtpi2ps  xmm6,mm4			; xmm6=(CD**)
+	cvtpi2ps  xmm7,mm1			; xmm7=(EF**)
+
+	movlhps   xmm0,xmm1			; xmm0=(0123)
+	movlhps   xmm2,xmm3			; xmm2=(4567)
+	movlhps   xmm4,xmm5			; xmm4=(89AB)
+	movlhps   xmm6,xmm7			; xmm6=(CDEF)
+
+	movaps	XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm2
+	movaps	XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm4
+	movaps	XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm6
+
+	add	esi, byte 2*SIZEOF_JSAMPROW
+	add	edi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT
+	dec	ecx
+	jnz	near .convloop
+
+	emms		; empty MMX state
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	pop	ebp
+	ret
+
+
+; --------------------------------------------------------------------------
+;
+; Quantize/descale the coefficients, and store into coef_block
+;
+; GLOBAL(void)
+; jsimd_quantize_float_sse (JCOEFPTR coef_block, FAST_FLOAT * divisors,
+;                           FAST_FLOAT * workspace);
+;
+
+%define coef_block	ebp+8		; JCOEFPTR coef_block
+%define divisors	ebp+12		; FAST_FLOAT * divisors
+%define workspace	ebp+16		; FAST_FLOAT * workspace
+
+	align	16
+	global	EXTN(jsimd_quantize_float_sse)
+
+EXTN(jsimd_quantize_float_sse):
+	push	ebp
+	mov	ebp,esp
+;	push	ebx		; unused
+;	push	ecx		; unused
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	mov	esi, POINTER [workspace]
+	mov	edx, POINTER [divisors]
+	mov	edi, JCOEFPTR [coef_block]
+	mov	eax, DCTSIZE2/16
+	alignx	16,7
+.quantloop:
+	movaps	xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(0,1,esi,SIZEOF_FAST_FLOAT)]
+	mulps	xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
+	mulps	xmm1, XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(1,1,esi,SIZEOF_FAST_FLOAT)]
+	mulps	xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
+	mulps	xmm3, XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)]
+
+	movhlps  xmm4,xmm0
+	movhlps  xmm5,xmm1
+
+	cvtps2pi mm0,xmm0
+	cvtps2pi mm1,xmm1
+	cvtps2pi mm4,xmm4
+	cvtps2pi mm5,xmm5
+
+	movhlps  xmm6,xmm2
+	movhlps  xmm7,xmm3
+
+	cvtps2pi mm2,xmm2
+	cvtps2pi mm3,xmm3
+	cvtps2pi mm6,xmm6
+	cvtps2pi mm7,xmm7
+
+	packssdw mm0,mm4
+	packssdw mm1,mm5
+	packssdw mm2,mm6
+	packssdw mm3,mm7
+
+	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0
+	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm1
+	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm2
+	movq	MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm3
+
+	add	esi, byte 16*SIZEOF_FAST_FLOAT
+	add	edx, byte 16*SIZEOF_FAST_FLOAT
+	add	edi, byte 16*SIZEOF_JCOEF
+	dec	eax
+	jnz	short .quantloop
+
+	emms		; empty MMX state
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; unused
+;	pop	ebx		; unused
+	pop	ebp
+	ret
+
diff --git a/common/jpeg/simd/jdct.inc b/common/jpeg/simd/jdct.inc
index e53f66e..cc62704 100644
--- a/common/jpeg/simd/jdct.inc
+++ b/common/jpeg/simd/jdct.inc
@@ -23,5 +23,6 @@
 
 %define DWBLOCK(m,n,b,s)	((b)+(m)*DCTSIZE*(s)+(n)*SIZEOF_DWORD)
 %define MMBLOCK(m,n,b,s)	((b)+(m)*DCTSIZE*(s)+(n)*SIZEOF_MMWORD)
+%define XMMBLOCK(m,n,b,s)	((b)+(m)*DCTSIZE*(s)+(n)*SIZEOF_XMMWORD)
 
 ; --------------------------------------------------------------------------
diff --git a/common/jpeg/simd/jfsseflt.asm b/common/jpeg/simd/jfsseflt.asm
new file mode 100644
index 0000000..6469f3c
--- /dev/null
+++ b/common/jpeg/simd/jfsseflt.asm
@@ -0,0 +1,367 @@
+;
+; jfsseflt.asm - floating-point FDCT (SSE)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a floating-point implementation of the forward DCT
+; (Discrete Cosine Transform). The following code is based directly on
+; the IJG's original jfdctflt.c; see the jfdctflt.c for more details.
+;
+; [TAB8]
+
+%include "simd/jsimdext.inc"
+%include "simd/jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%macro	unpcklps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
+	shufps	%1,%2,0x44
+%endmacro
+
+%macro	unpckhps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
+	shufps	%1,%2,0xEE
+%endmacro
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_fdct_float_sse)
+
+EXTN(jconst_fdct_float_sse):
+
+PD_0_382	times 4 dd  0.382683432365089771728460
+PD_0_707	times 4 dd  0.707106781186547524400844
+PD_0_541	times 4 dd  0.541196100146196984399723
+PD_1_306	times 4 dd  1.306562964876376527856643
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform the forward DCT on one block of samples.
+;
+; GLOBAL(void)
+; jsimd_fdct_float_sse (FAST_FLOAT * data)
+;
+
+%define data(b)		(b)+8		; FAST_FLOAT * data
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		2
+
+	align	16
+	global	EXTN(jsimd_fdct_float_sse)
+
+EXTN(jsimd_fdct_float_sse):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [wk(0)]
+	pushpic	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+;	push	esi		; unused
+;	push	edi		; unused
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process rows.
+
+	mov	edx, POINTER [data(eax)]	; (FAST_FLOAT *)
+	mov	ecx, DCTSIZE/4
+	alignx	16,7
+.rowloop:
+
+	movaps	xmm0, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm2, XMMWORD [XMMBLOCK(2,1,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(3,1,edx,SIZEOF_FAST_FLOAT)]
+
+	; xmm0=(20 21 22 23), xmm2=(24 25 26 27)
+	; xmm1=(30 31 32 33), xmm3=(34 35 36 37)
+
+	movaps   xmm4,xmm0		; transpose coefficients(phase 1)
+	unpcklps xmm0,xmm1		; xmm0=(20 30 21 31)
+	unpckhps xmm4,xmm1		; xmm4=(22 32 23 33)
+	movaps   xmm5,xmm2		; transpose coefficients(phase 1)
+	unpcklps xmm2,xmm3		; xmm2=(24 34 25 35)
+	unpckhps xmm5,xmm3		; xmm5=(26 36 27 37)
+
+	movaps	xmm6, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm7, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)]
+
+	; xmm6=(00 01 02 03), xmm1=(04 05 06 07)
+	; xmm7=(10 11 12 13), xmm3=(14 15 16 17)
+
+	movaps	XMMWORD [wk(0)], xmm4	; wk(0)=(22 32 23 33)
+	movaps	XMMWORD [wk(1)], xmm2	; wk(1)=(24 34 25 35)
+
+	movaps   xmm4,xmm6		; transpose coefficients(phase 1)
+	unpcklps xmm6,xmm7		; xmm6=(00 10 01 11)
+	unpckhps xmm4,xmm7		; xmm4=(02 12 03 13)
+	movaps   xmm2,xmm1		; transpose coefficients(phase 1)
+	unpcklps xmm1,xmm3		; xmm1=(04 14 05 15)
+	unpckhps xmm2,xmm3		; xmm2=(06 16 07 17)
+
+	movaps    xmm7,xmm6		; transpose coefficients(phase 2)
+	unpcklps2 xmm6,xmm0		; xmm6=(00 10 20 30)=data0
+	unpckhps2 xmm7,xmm0		; xmm7=(01 11 21 31)=data1
+	movaps    xmm3,xmm2		; transpose coefficients(phase 2)
+	unpcklps2 xmm2,xmm5		; xmm2=(06 16 26 36)=data6
+	unpckhps2 xmm3,xmm5		; xmm3=(07 17 27 37)=data7
+
+	movaps	xmm0,xmm7
+	movaps	xmm5,xmm6
+	subps	xmm7,xmm2		; xmm7=data1-data6=tmp6
+	subps	xmm6,xmm3		; xmm6=data0-data7=tmp7
+	addps	xmm0,xmm2		; xmm0=data1+data6=tmp1
+	addps	xmm5,xmm3		; xmm5=data0+data7=tmp0
+
+	movaps	xmm2, XMMWORD [wk(0)]	; xmm2=(22 32 23 33)
+	movaps	xmm3, XMMWORD [wk(1)]	; xmm3=(24 34 25 35)
+	movaps	XMMWORD [wk(0)], xmm7	; wk(0)=tmp6
+	movaps	XMMWORD [wk(1)], xmm6	; wk(1)=tmp7
+
+	movaps    xmm7,xmm4		; transpose coefficients(phase 2)
+	unpcklps2 xmm4,xmm2		; xmm4=(02 12 22 32)=data2
+	unpckhps2 xmm7,xmm2		; xmm7=(03 13 23 33)=data3
+	movaps    xmm6,xmm1		; transpose coefficients(phase 2)
+	unpcklps2 xmm1,xmm3		; xmm1=(04 14 24 34)=data4
+	unpckhps2 xmm6,xmm3		; xmm6=(05 15 25 35)=data5
+
+	movaps	xmm2,xmm7
+	movaps	xmm3,xmm4
+	addps	xmm7,xmm1		; xmm7=data3+data4=tmp3
+	addps	xmm4,xmm6		; xmm4=data2+data5=tmp2
+	subps	xmm2,xmm1		; xmm2=data3-data4=tmp4
+	subps	xmm3,xmm6		; xmm3=data2-data5=tmp5
+
+	; -- Even part
+
+	movaps	xmm1,xmm5
+	movaps	xmm6,xmm0
+	subps	xmm5,xmm7		; xmm5=tmp13
+	subps	xmm0,xmm4		; xmm0=tmp12
+	addps	xmm1,xmm7		; xmm1=tmp10
+	addps	xmm6,xmm4		; xmm6=tmp11
+
+	addps	xmm0,xmm5
+	mulps	xmm0,[GOTOFF(ebx,PD_0_707)] ; xmm0=z1
+
+	movaps	xmm7,xmm1
+	movaps	xmm4,xmm5
+	subps	xmm1,xmm6		; xmm1=data4
+	subps	xmm5,xmm0		; xmm5=data6
+	addps	xmm7,xmm6		; xmm7=data0
+	addps	xmm4,xmm0		; xmm4=data2
+
+	movaps	XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(2,1,edx,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], xmm7
+	movaps	XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)], xmm4
+
+	; -- Odd part
+
+	movaps	xmm6, XMMWORD [wk(0)]	; xmm6=tmp6
+	movaps	xmm0, XMMWORD [wk(1)]	; xmm0=tmp7
+
+	addps	xmm2,xmm3		; xmm2=tmp10
+	addps	xmm3,xmm6		; xmm3=tmp11
+	addps	xmm6,xmm0		; xmm6=tmp12, xmm0=tmp7
+
+	mulps	xmm3,[GOTOFF(ebx,PD_0_707)] ; xmm3=z3
+
+	movaps	xmm1,xmm2		; xmm1=tmp10
+	subps	xmm2,xmm6
+	mulps	xmm2,[GOTOFF(ebx,PD_0_382)] ; xmm2=z5
+	mulps	xmm1,[GOTOFF(ebx,PD_0_541)] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
+	mulps	xmm6,[GOTOFF(ebx,PD_1_306)] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
+	addps	xmm1,xmm2		; xmm1=z2
+	addps	xmm6,xmm2		; xmm6=z4
+
+	movaps	xmm5,xmm0
+	subps	xmm0,xmm3		; xmm0=z13
+	addps	xmm5,xmm3		; xmm5=z11
+
+	movaps	xmm7,xmm0
+	movaps	xmm4,xmm5
+	subps	xmm0,xmm1		; xmm0=data3
+	subps	xmm5,xmm6		; xmm5=data7
+	addps	xmm7,xmm1		; xmm7=data5
+	addps	xmm4,xmm6		; xmm4=data1
+
+	movaps	XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(3,1,edx,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)], xmm7
+	movaps	XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], xmm4
+
+	add	edx, 4*DCTSIZE*SIZEOF_FAST_FLOAT
+	dec	ecx
+	jnz	near .rowloop
+
+	; ---- Pass 2: process columns.
+
+	mov	edx, POINTER [data(eax)]	; (FAST_FLOAT *)
+	mov	ecx, DCTSIZE/4
+	alignx	16,7
+.columnloop:
+
+	movaps	xmm0, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm2, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)]
+
+	; xmm0=(02 12 22 32), xmm2=(42 52 62 72)
+	; xmm1=(03 13 23 33), xmm3=(43 53 63 73)
+
+	movaps   xmm4,xmm0		; transpose coefficients(phase 1)
+	unpcklps xmm0,xmm1		; xmm0=(02 03 12 13)
+	unpckhps xmm4,xmm1		; xmm4=(22 23 32 33)
+	movaps   xmm5,xmm2		; transpose coefficients(phase 1)
+	unpcklps xmm2,xmm3		; xmm2=(42 43 52 53)
+	unpckhps xmm5,xmm3		; xmm5=(62 63 72 73)
+
+	movaps	xmm6, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm7, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)]
+
+	; xmm6=(00 10 20 30), xmm1=(40 50 60 70)
+	; xmm7=(01 11 21 31), xmm3=(41 51 61 71)
+
+	movaps	XMMWORD [wk(0)], xmm4	; wk(0)=(22 23 32 33)
+	movaps	XMMWORD [wk(1)], xmm2	; wk(1)=(42 43 52 53)
+
+	movaps   xmm4,xmm6		; transpose coefficients(phase 1)
+	unpcklps xmm6,xmm7		; xmm6=(00 01 10 11)
+	unpckhps xmm4,xmm7		; xmm4=(20 21 30 31)
+	movaps   xmm2,xmm1		; transpose coefficients(phase 1)
+	unpcklps xmm1,xmm3		; xmm1=(40 41 50 51)
+	unpckhps xmm2,xmm3		; xmm2=(60 61 70 71)
+
+	movaps    xmm7,xmm6		; transpose coefficients(phase 2)
+	unpcklps2 xmm6,xmm0		; xmm6=(00 01 02 03)=data0
+	unpckhps2 xmm7,xmm0		; xmm7=(10 11 12 13)=data1
+	movaps    xmm3,xmm2		; transpose coefficients(phase 2)
+	unpcklps2 xmm2,xmm5		; xmm2=(60 61 62 63)=data6
+	unpckhps2 xmm3,xmm5		; xmm3=(70 71 72 73)=data7
+
+	movaps	xmm0,xmm7
+	movaps	xmm5,xmm6
+	subps	xmm7,xmm2		; xmm7=data1-data6=tmp6
+	subps	xmm6,xmm3		; xmm6=data0-data7=tmp7
+	addps	xmm0,xmm2		; xmm0=data1+data6=tmp1
+	addps	xmm5,xmm3		; xmm5=data0+data7=tmp0
+
+	movaps	xmm2, XMMWORD [wk(0)]	; xmm2=(22 23 32 33)
+	movaps	xmm3, XMMWORD [wk(1)]	; xmm3=(42 43 52 53)
+	movaps	XMMWORD [wk(0)], xmm7	; wk(0)=tmp6
+	movaps	XMMWORD [wk(1)], xmm6	; wk(1)=tmp7
+
+	movaps    xmm7,xmm4		; transpose coefficients(phase 2)
+	unpcklps2 xmm4,xmm2		; xmm4=(20 21 22 23)=data2
+	unpckhps2 xmm7,xmm2		; xmm7=(30 31 32 33)=data3
+	movaps    xmm6,xmm1		; transpose coefficients(phase 2)
+	unpcklps2 xmm1,xmm3		; xmm1=(40 41 42 43)=data4
+	unpckhps2 xmm6,xmm3		; xmm6=(50 51 52 53)=data5
+
+	movaps	xmm2,xmm7
+	movaps	xmm3,xmm4
+	addps	xmm7,xmm1		; xmm7=data3+data4=tmp3
+	addps	xmm4,xmm6		; xmm4=data2+data5=tmp2
+	subps	xmm2,xmm1		; xmm2=data3-data4=tmp4
+	subps	xmm3,xmm6		; xmm3=data2-data5=tmp5
+
+	; -- Even part
+
+	movaps	xmm1,xmm5
+	movaps	xmm6,xmm0
+	subps	xmm5,xmm7		; xmm5=tmp13
+	subps	xmm0,xmm4		; xmm0=tmp12
+	addps	xmm1,xmm7		; xmm1=tmp10
+	addps	xmm6,xmm4		; xmm6=tmp11
+
+	addps	xmm0,xmm5
+	mulps	xmm0,[GOTOFF(ebx,PD_0_707)] ; xmm0=z1
+
+	movaps	xmm7,xmm1
+	movaps	xmm4,xmm5
+	subps	xmm1,xmm6		; xmm1=data4
+	subps	xmm5,xmm0		; xmm5=data6
+	addps	xmm7,xmm6		; xmm7=data0
+	addps	xmm4,xmm0		; xmm4=data2
+
+	movaps	XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], xmm7
+	movaps	XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)], xmm4
+
+	; -- Odd part
+
+	movaps	xmm6, XMMWORD [wk(0)]	; xmm6=tmp6
+	movaps	xmm0, XMMWORD [wk(1)]	; xmm0=tmp7
+
+	addps	xmm2,xmm3		; xmm2=tmp10
+	addps	xmm3,xmm6		; xmm3=tmp11
+	addps	xmm6,xmm0		; xmm6=tmp12, xmm0=tmp7
+
+	mulps	xmm3,[GOTOFF(ebx,PD_0_707)] ; xmm3=z3
+
+	movaps	xmm1,xmm2		; xmm1=tmp10
+	subps	xmm2,xmm6
+	mulps	xmm2,[GOTOFF(ebx,PD_0_382)] ; xmm2=z5
+	mulps	xmm1,[GOTOFF(ebx,PD_0_541)] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
+	mulps	xmm6,[GOTOFF(ebx,PD_1_306)] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
+	addps	xmm1,xmm2		; xmm1=z2
+	addps	xmm6,xmm2		; xmm6=z4
+
+	movaps	xmm5,xmm0
+	subps	xmm0,xmm3		; xmm0=z13
+	addps	xmm5,xmm3		; xmm5=z11
+
+	movaps	xmm7,xmm0
+	movaps	xmm4,xmm5
+	subps	xmm0,xmm1		; xmm0=data3
+	subps	xmm5,xmm6		; xmm5=data7
+	addps	xmm7,xmm1		; xmm7=data5
+	addps	xmm4,xmm6		; xmm4=data1
+
+	movaps	XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)], xmm7
+	movaps	XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], xmm4
+
+	add	edx, byte 4*SIZEOF_FAST_FLOAT
+	dec	ecx
+	jnz	near .columnloop
+
+;	pop	edi		; unused
+;	pop	esi		; unused
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	poppic	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
diff --git a/common/jpeg/simd/jisseflt.asm b/common/jpeg/simd/jisseflt.asm
new file mode 100644
index 0000000..6de93ad
--- /dev/null
+++ b/common/jpeg/simd/jisseflt.asm
@@ -0,0 +1,569 @@
+;
+; jisseflt.asm - floating-point IDCT (SSE & MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; x86 SIMD extension for IJG JPEG library
+; Copyright (C) 1999-2006, MIYASAKA Masaru.
+; For conditions of distribution and use, see copyright notice in jsimdext.inc
+;
+; This file should be assembled with NASM (Netwide Assembler),
+; can *not* be assembled with Microsoft's MASM or any compatible
+; assembler (including Borland's Turbo Assembler).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; This file contains a floating-point implementation of the inverse DCT
+; (Discrete Cosine Transform). The following code is based directly on
+; the IJG's original jidctflt.c; see the jidctflt.c for more details.
+;
+; [TAB8]
+
+%include "simd/jsimdext.inc"
+%include "simd/jdct.inc"
+
+; --------------------------------------------------------------------------
+
+%macro	unpcklps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
+	shufps	%1,%2,0x44
+%endmacro
+
+%macro	unpckhps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
+	shufps	%1,%2,0xEE
+%endmacro
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_CONST
+
+	alignz	16
+	global	EXTN(jconst_idct_float_sse)
+
+EXTN(jconst_idct_float_sse):
+
+PD_1_414	times 4 dd  1.414213562373095048801689
+PD_1_847	times 4 dd  1.847759065022573512256366
+PD_1_082	times 4 dd  1.082392200292393968799446
+PD_M2_613	times 4 dd -2.613125929752753055713286
+PD_0_125	times 4 dd  0.125	; 1/8
+PB_CENTERJSAMP	times 8 db  CENTERJSAMPLE
+
+	alignz	16
+
+; --------------------------------------------------------------------------
+	SECTION	SEG_TEXT
+	BITS	32
+;
+; Perform dequantization and inverse DCT on one block of coefficients.
+;
+; GLOBAL(void)
+; jsimd_idct_float_sse (void * dct_table, JCOEFPTR coef_block,
+;                       JSAMPARRAY output_buf, JDIMENSION output_col)
+;
+
+%define dct_table(b)	(b)+8			; void * dct_table
+%define coef_block(b)	(b)+12		; JCOEFPTR coef_block
+%define output_buf(b)	(b)+16		; JSAMPARRAY output_buf
+%define output_col(b)	(b)+20		; JDIMENSION output_col
+
+%define original_ebp	ebp+0
+%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
+%define WK_NUM		2
+%define workspace	wk(0)-DCTSIZE2*SIZEOF_FAST_FLOAT
+					; FAST_FLOAT workspace[DCTSIZE2]
+
+	align	16
+	global	EXTN(jsimd_idct_float_sse)
+
+EXTN(jsimd_idct_float_sse):
+	push	ebp
+	mov	eax,esp				; eax = original ebp
+	sub	esp, byte 4
+	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
+	mov	[esp],eax
+	mov	ebp,esp				; ebp = aligned ebp
+	lea	esp, [workspace]
+	push	ebx
+;	push	ecx		; need not be preserved
+;	push	edx		; need not be preserved
+	push	esi
+	push	edi
+
+	get_GOT	ebx		; get GOT address
+
+	; ---- Pass 1: process columns from input, store into work array.
+
+;	mov	eax, [original_ebp]
+	mov	edx, POINTER [dct_table(eax)]	; quantptr
+	mov	esi, JCOEFPTR [coef_block(eax)]		; inptr
+	lea	edi, [workspace]			; FAST_FLOAT * wsptr
+	mov	ecx, DCTSIZE/4				; ctr
+	alignx	16,7
+.columnloop:
+%ifndef NO_ZERO_COLUMN_TEST_FLOAT_SSE
+	mov	eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	or	eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	jnz	near .columnDCT
+
+	movq	mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	por	mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	por	mm1, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	por	mm0, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	por	mm1, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+	por	mm0, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+	por	mm1,mm0
+	packsswb mm1,mm1
+	movd	eax,mm1
+	test	eax,eax
+	jnz	short .columnDCT
+
+	; -- AC terms all zero
+
+	movq      mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+
+	punpckhwd mm1,mm0			; mm1=(** 02 ** 03)
+	punpcklwd mm0,mm0			; mm0=(00 00 01 01)
+	psrad     mm1,(DWORD_BIT-WORD_BIT)	; mm1=in0H=(02 03)
+	psrad     mm0,(DWORD_BIT-WORD_BIT)	; mm0=in0L=(00 01)
+	cvtpi2ps  xmm3,mm1			; xmm3=(02 03 ** **)
+	cvtpi2ps  xmm0,mm0			; xmm0=(00 01 ** **)
+	movlhps   xmm0,xmm3			; xmm0=in0=(00 01 02 03)
+
+	mulps	xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movaps	xmm1,xmm0
+	movaps	xmm2,xmm0
+	movaps	xmm3,xmm0
+
+	shufps	xmm0,xmm0,0x00			; xmm0=(00 00 00 00)
+	shufps	xmm1,xmm1,0x55			; xmm1=(01 01 01 01)
+	shufps	xmm2,xmm2,0xAA			; xmm2=(02 02 02 02)
+	shufps	xmm3,xmm3,0xFF			; xmm3=(03 03 03 03)
+
+	movaps	XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm0
+	movaps	XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm2
+	movaps	XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm2
+	movaps	XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm3
+	movaps	XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3
+	jmp	near .nextcolumn
+	alignx	16,7
+%endif
+.columnDCT:
+
+	; -- Even part
+
+	movq      mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
+	movq      mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
+	movq      mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
+	movq      mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
+
+	punpckhwd mm4,mm0			; mm4=(** 02 ** 03)
+	punpcklwd mm0,mm0			; mm0=(00 00 01 01)
+	punpckhwd mm5,mm1			; mm5=(** 22 ** 23)
+	punpcklwd mm1,mm1			; mm1=(20 20 21 21)
+
+	psrad     mm4,(DWORD_BIT-WORD_BIT)	; mm4=in0H=(02 03)
+	psrad     mm0,(DWORD_BIT-WORD_BIT)	; mm0=in0L=(00 01)
+	cvtpi2ps  xmm4,mm4			; xmm4=(02 03 ** **)
+	cvtpi2ps  xmm0,mm0			; xmm0=(00 01 ** **)
+	psrad     mm5,(DWORD_BIT-WORD_BIT)	; mm5=in2H=(22 23)
+	psrad     mm1,(DWORD_BIT-WORD_BIT)	; mm1=in2L=(20 21)
+	cvtpi2ps  xmm5,mm5			; xmm5=(22 23 ** **)
+	cvtpi2ps  xmm1,mm1			; xmm1=(20 21 ** **)
+
+	punpckhwd mm6,mm2			; mm6=(** 42 ** 43)
+	punpcklwd mm2,mm2			; mm2=(40 40 41 41)
+	punpckhwd mm7,mm3			; mm7=(** 62 ** 63)
+	punpcklwd mm3,mm3			; mm3=(60 60 61 61)
+
+	psrad     mm6,(DWORD_BIT-WORD_BIT)	; mm6=in4H=(42 43)
+	psrad     mm2,(DWORD_BIT-WORD_BIT)	; mm2=in4L=(40 41)
+	cvtpi2ps  xmm6,mm6			; xmm6=(42 43 ** **)
+	cvtpi2ps  xmm2,mm2			; xmm2=(40 41 ** **)
+	psrad     mm7,(DWORD_BIT-WORD_BIT)	; mm7=in6H=(62 63)
+	psrad     mm3,(DWORD_BIT-WORD_BIT)	; mm3=in6L=(60 61)
+	cvtpi2ps  xmm7,mm7			; xmm7=(62 63 ** **)
+	cvtpi2ps  xmm3,mm3			; xmm3=(60 61 ** **)
+
+	movlhps   xmm0,xmm4			; xmm0=in0=(00 01 02 03)
+	movlhps   xmm1,xmm5			; xmm1=in2=(20 21 22 23)
+	mulps     xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm1, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movlhps   xmm2,xmm6			; xmm2=in4=(40 41 42 43)
+	movlhps   xmm3,xmm7			; xmm3=in6=(60 61 62 63)
+	mulps     xmm2, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm3, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movaps	xmm4,xmm0
+	movaps	xmm5,xmm1
+	subps	xmm0,xmm2		; xmm0=tmp11
+	subps	xmm1,xmm3
+	addps	xmm4,xmm2		; xmm4=tmp10
+	addps	xmm5,xmm3		; xmm5=tmp13
+
+	mulps	xmm1,[GOTOFF(ebx,PD_1_414)]
+	subps	xmm1,xmm5		; xmm1=tmp12
+
+	movaps	xmm6,xmm4
+	movaps	xmm7,xmm0
+	subps	xmm4,xmm5		; xmm4=tmp3
+	subps	xmm0,xmm1		; xmm0=tmp2
+	addps	xmm6,xmm5		; xmm6=tmp0
+	addps	xmm7,xmm1		; xmm7=tmp1
+
+	movaps	XMMWORD [wk(1)], xmm4	; tmp3
+	movaps	XMMWORD [wk(0)], xmm0	; tmp2
+
+	; -- Odd part
+
+	movq      mm4, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
+	movq      mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
+	movq      mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
+	movq      mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
+
+	punpckhwd mm6,mm4			; mm6=(** 12 ** 13)
+	punpcklwd mm4,mm4			; mm4=(10 10 11 11)
+	punpckhwd mm2,mm0			; mm2=(** 32 ** 33)
+	punpcklwd mm0,mm0			; mm0=(30 30 31 31)
+
+	psrad     mm6,(DWORD_BIT-WORD_BIT)	; mm6=in1H=(12 13)
+	psrad     mm4,(DWORD_BIT-WORD_BIT)	; mm4=in1L=(10 11)
+	cvtpi2ps  xmm4,mm6			; xmm4=(12 13 ** **)
+	cvtpi2ps  xmm2,mm4			; xmm2=(10 11 ** **)
+	psrad     mm2,(DWORD_BIT-WORD_BIT)	; mm2=in3H=(32 33)
+	psrad     mm0,(DWORD_BIT-WORD_BIT)	; mm0=in3L=(30 31)
+	cvtpi2ps  xmm0,mm2			; xmm0=(32 33 ** **)
+	cvtpi2ps  xmm3,mm0			; xmm3=(30 31 ** **)
+
+	punpckhwd mm7,mm5			; mm7=(** 52 ** 53)
+	punpcklwd mm5,mm5			; mm5=(50 50 51 51)
+	punpckhwd mm3,mm1			; mm3=(** 72 ** 73)
+	punpcklwd mm1,mm1			; mm1=(70 70 71 71)
+
+	movlhps   xmm2,xmm4			; xmm2=in1=(10 11 12 13)
+	movlhps   xmm3,xmm0			; xmm3=in3=(30 31 32 33)
+
+	psrad     mm7,(DWORD_BIT-WORD_BIT)	; mm7=in5H=(52 53)
+	psrad     mm5,(DWORD_BIT-WORD_BIT)	; mm5=in5L=(50 51)
+	cvtpi2ps  xmm4,mm7			; xmm4=(52 53 ** **)
+	cvtpi2ps  xmm5,mm5			; xmm5=(50 51 ** **)
+	psrad     mm3,(DWORD_BIT-WORD_BIT)	; mm3=in7H=(72 73)
+	psrad     mm1,(DWORD_BIT-WORD_BIT)	; mm1=in7L=(70 71)
+	cvtpi2ps  xmm0,mm3			; xmm0=(72 73 ** **)
+	cvtpi2ps  xmm1,mm1			; xmm1=(70 71 ** **)
+
+	mulps     xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movlhps   xmm5,xmm4			; xmm5=in5=(50 51 52 53)
+	movlhps   xmm1,xmm0			; xmm1=in7=(70 71 72 73)
+	mulps     xmm5, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+	mulps     xmm1, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
+
+	movaps	xmm4,xmm2
+	movaps	xmm0,xmm5
+	addps	xmm2,xmm1		; xmm2=z11
+	addps	xmm5,xmm3		; xmm5=z13
+	subps	xmm4,xmm1		; xmm4=z12
+	subps	xmm0,xmm3		; xmm0=z10
+
+	movaps	xmm1,xmm2
+	subps	xmm2,xmm5
+	addps	xmm1,xmm5		; xmm1=tmp7
+
+	mulps	xmm2,[GOTOFF(ebx,PD_1_414)]	; xmm2=tmp11
+
+	movaps	xmm3,xmm0
+	addps	xmm0,xmm4
+	mulps	xmm0,[GOTOFF(ebx,PD_1_847)]	; xmm0=z5
+	mulps	xmm3,[GOTOFF(ebx,PD_M2_613)]	; xmm3=(z10 * -2.613125930)
+	mulps	xmm4,[GOTOFF(ebx,PD_1_082)]	; xmm4=(z12 * 1.082392200)
+	addps	xmm3,xmm0		; xmm3=tmp12
+	subps	xmm4,xmm0		; xmm4=tmp10
+
+	; -- Final output stage
+
+	subps	xmm3,xmm1		; xmm3=tmp6
+	movaps	xmm5,xmm6
+	movaps	xmm0,xmm7
+	addps	xmm6,xmm1		; xmm6=data0=(00 01 02 03)
+	addps	xmm7,xmm3		; xmm7=data1=(10 11 12 13)
+	subps	xmm5,xmm1		; xmm5=data7=(70 71 72 73)
+	subps	xmm0,xmm3		; xmm0=data6=(60 61 62 63)
+	subps	xmm2,xmm3		; xmm2=tmp5
+
+	movaps    xmm1,xmm6		; transpose coefficients(phase 1)
+	unpcklps  xmm6,xmm7		; xmm6=(00 10 01 11)
+	unpckhps  xmm1,xmm7		; xmm1=(02 12 03 13)
+	movaps    xmm3,xmm0		; transpose coefficients(phase 1)
+	unpcklps  xmm0,xmm5		; xmm0=(60 70 61 71)
+	unpckhps  xmm3,xmm5		; xmm3=(62 72 63 73)
+
+	movaps	xmm7, XMMWORD [wk(0)]	; xmm7=tmp2
+	movaps	xmm5, XMMWORD [wk(1)]	; xmm5=tmp3
+
+	movaps	XMMWORD [wk(0)], xmm0	; wk(0)=(60 70 61 71)
+	movaps	XMMWORD [wk(1)], xmm3	; wk(1)=(62 72 63 73)
+
+	addps	xmm4,xmm2		; xmm4=tmp4
+	movaps	xmm0,xmm7
+	movaps	xmm3,xmm5
+	addps	xmm7,xmm2		; xmm7=data2=(20 21 22 23)
+	addps	xmm5,xmm4		; xmm5=data4=(40 41 42 43)
+	subps	xmm0,xmm2		; xmm0=data5=(50 51 52 53)
+	subps	xmm3,xmm4		; xmm3=data3=(30 31 32 33)
+
+	movaps    xmm2,xmm7		; transpose coefficients(phase 1)
+	unpcklps  xmm7,xmm3		; xmm7=(20 30 21 31)
+	unpckhps  xmm2,xmm3		; xmm2=(22 32 23 33)
+	movaps    xmm4,xmm5		; transpose coefficients(phase 1)
+	unpcklps  xmm5,xmm0		; xmm5=(40 50 41 51)
+	unpckhps  xmm4,xmm0		; xmm4=(42 52 43 53)
+
+	movaps    xmm3,xmm6		; transpose coefficients(phase 2)
+	unpcklps2 xmm6,xmm7		; xmm6=(00 10 20 30)
+	unpckhps2 xmm3,xmm7		; xmm3=(01 11 21 31)
+	movaps    xmm0,xmm1		; transpose coefficients(phase 2)
+	unpcklps2 xmm1,xmm2		; xmm1=(02 12 22 32)
+	unpckhps2 xmm0,xmm2		; xmm0=(03 13 23 33)
+
+	movaps	xmm7, XMMWORD [wk(0)]	; xmm7=(60 70 61 71)
+	movaps	xmm2, XMMWORD [wk(1)]	; xmm2=(62 72 63 73)
+
+	movaps	XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm6
+	movaps	XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm3
+	movaps	XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm1
+	movaps	XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm0
+
+	movaps    xmm6,xmm5		; transpose coefficients(phase 2)
+	unpcklps2 xmm5,xmm7		; xmm5=(40 50 60 70)
+	unpckhps2 xmm6,xmm7		; xmm6=(41 51 61 71)
+	movaps    xmm3,xmm4		; transpose coefficients(phase 2)
+	unpcklps2 xmm4,xmm2		; xmm4=(42 52 62 72)
+	unpckhps2 xmm3,xmm2		; xmm3=(43 53 63 73)
+
+	movaps	XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm5
+	movaps	XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm6
+	movaps	XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm4
+	movaps	XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3
+
+.nextcolumn:
+	add	esi, byte 4*SIZEOF_JCOEF		; coef_block
+	add	edx, byte 4*SIZEOF_FLOAT_MULT_TYPE	; quantptr
+	add	edi,      4*DCTSIZE*SIZEOF_FAST_FLOAT	; wsptr
+	dec	ecx					; ctr
+	jnz	near .columnloop
+
+	; -- Prefetch the next coefficient block
+
+	prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 0*32]
+	prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 1*32]
+	prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 2*32]
+	prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 3*32]
+
+	; ---- Pass 2: process rows from work array, store into output array.
+
+	mov	eax, [original_ebp]
+	lea	esi, [workspace]			; FAST_FLOAT * wsptr
+	mov	edi, JSAMPARRAY [output_buf(eax)]	; (JSAMPROW *)
+	mov	eax, JDIMENSION [output_col(eax)]
+	mov	ecx, DCTSIZE/4				; ctr
+	alignx	16,7
+.rowloop:
+
+	; -- Even part
+
+	movaps	xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm2, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_FAST_FLOAT)]
+
+	movaps	xmm4,xmm0
+	movaps	xmm5,xmm1
+	subps	xmm0,xmm2		; xmm0=tmp11
+	subps	xmm1,xmm3
+	addps	xmm4,xmm2		; xmm4=tmp10
+	addps	xmm5,xmm3		; xmm5=tmp13
+
+	mulps	xmm1,[GOTOFF(ebx,PD_1_414)]
+	subps	xmm1,xmm5		; xmm1=tmp12
+
+	movaps	xmm6,xmm4
+	movaps	xmm7,xmm0
+	subps	xmm4,xmm5		; xmm4=tmp3
+	subps	xmm0,xmm1		; xmm0=tmp2
+	addps	xmm6,xmm5		; xmm6=tmp0
+	addps	xmm7,xmm1		; xmm7=tmp1
+
+	movaps	XMMWORD [wk(1)], xmm4	; tmp3
+	movaps	XMMWORD [wk(0)], xmm0	; tmp2
+
+	; -- Odd part
+
+	movaps	xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm3, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm5, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_FAST_FLOAT)]
+	movaps	xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_FAST_FLOAT)]
+
+	movaps	xmm4,xmm2
+	movaps	xmm0,xmm5
+	addps	xmm2,xmm1		; xmm2=z11
+	addps	xmm5,xmm3		; xmm5=z13
+	subps	xmm4,xmm1		; xmm4=z12
+	subps	xmm0,xmm3		; xmm0=z10
+
+	movaps	xmm1,xmm2
+	subps	xmm2,xmm5
+	addps	xmm1,xmm5		; xmm1=tmp7
+
+	mulps	xmm2,[GOTOFF(ebx,PD_1_414)]	; xmm2=tmp11
+
+	movaps	xmm3,xmm0
+	addps	xmm0,xmm4
+	mulps	xmm0,[GOTOFF(ebx,PD_1_847)]	; xmm0=z5
+	mulps	xmm3,[GOTOFF(ebx,PD_M2_613)]	; xmm3=(z10 * -2.613125930)
+	mulps	xmm4,[GOTOFF(ebx,PD_1_082)]	; xmm4=(z12 * 1.082392200)
+	addps	xmm3,xmm0		; xmm3=tmp12
+	subps	xmm4,xmm0		; xmm4=tmp10
+
+	; -- Final output stage
+
+	subps	xmm3,xmm1		; xmm3=tmp6
+	movaps	xmm5,xmm6
+	movaps	xmm0,xmm7
+	addps	xmm6,xmm1		; xmm6=data0=(00 10 20 30)
+	addps	xmm7,xmm3		; xmm7=data1=(01 11 21 31)
+	subps	xmm5,xmm1		; xmm5=data7=(07 17 27 37)
+	subps	xmm0,xmm3		; xmm0=data6=(06 16 26 36)
+	subps	xmm2,xmm3		; xmm2=tmp5
+
+	movaps	xmm1,[GOTOFF(ebx,PD_0_125)]	; xmm1=[PD_0_125]
+
+	mulps	xmm6,xmm1		; descale(1/8)
+	mulps	xmm7,xmm1		; descale(1/8)
+	mulps	xmm5,xmm1		; descale(1/8)
+	mulps	xmm0,xmm1		; descale(1/8)
+
+	movhlps   xmm3,xmm6
+	movhlps   xmm1,xmm7
+	cvtps2pi  mm0,xmm6		; round to int32, mm0=data0L=(00 10)
+	cvtps2pi  mm1,xmm7		; round to int32, mm1=data1L=(01 11)
+	cvtps2pi  mm2,xmm3		; round to int32, mm2=data0H=(20 30)
+	cvtps2pi  mm3,xmm1		; round to int32, mm3=data1H=(21 31)
+	packssdw  mm0,mm2		; mm0=data0=(00 10 20 30)
+	packssdw  mm1,mm3		; mm1=data1=(01 11 21 31)
+
+	movhlps   xmm6,xmm5
+	movhlps   xmm7,xmm0
+	cvtps2pi  mm4,xmm5		; round to int32, mm4=data7L=(07 17)
+	cvtps2pi  mm5,xmm0		; round to int32, mm5=data6L=(06 16)
+	cvtps2pi  mm6,xmm6		; round to int32, mm6=data7H=(27 37)
+	cvtps2pi  mm7,xmm7		; round to int32, mm7=data6H=(26 36)
+	packssdw  mm4,mm6		; mm4=data7=(07 17 27 37)
+	packssdw  mm5,mm7		; mm5=data6=(06 16 26 36)
+
+	packsswb  mm0,mm5		; mm0=(00 10 20 30 06 16 26 36)
+	packsswb  mm1,mm4		; mm1=(01 11 21 31 07 17 27 37)
+
+	movaps	xmm3, XMMWORD [wk(0)]	; xmm3=tmp2
+	movaps	xmm1, XMMWORD [wk(1)]	; xmm1=tmp3
+
+	movaps	xmm6,[GOTOFF(ebx,PD_0_125)]	; xmm6=[PD_0_125]
+
+	addps	xmm4,xmm2		; xmm4=tmp4
+	movaps	xmm5,xmm3
+	movaps	xmm0,xmm1
+	addps	xmm3,xmm2		; xmm3=data2=(02 12 22 32)
+	addps	xmm1,xmm4		; xmm1=data4=(04 14 24 34)
+	subps	xmm5,xmm2		; xmm5=data5=(05 15 25 35)
+	subps	xmm0,xmm4		; xmm0=data3=(03 13 23 33)
+
+	mulps	xmm3,xmm6		; descale(1/8)
+	mulps	xmm1,xmm6		; descale(1/8)
+	mulps	xmm5,xmm6		; descale(1/8)
+	mulps	xmm0,xmm6		; descale(1/8)
+
+	movhlps   xmm7,xmm3
+	movhlps   xmm2,xmm1
+	cvtps2pi  mm2,xmm3		; round to int32, mm2=data2L=(02 12)
+	cvtps2pi  mm3,xmm1		; round to int32, mm3=data4L=(04 14)
+	cvtps2pi  mm6,xmm7		; round to int32, mm6=data2H=(22 32)
+	cvtps2pi  mm7,xmm2		; round to int32, mm7=data4H=(24 34)
+	packssdw  mm2,mm6		; mm2=data2=(02 12 22 32)
+	packssdw  mm3,mm7		; mm3=data4=(04 14 24 34)
+
+	movhlps   xmm4,xmm5
+	movhlps   xmm6,xmm0
+	cvtps2pi  mm5,xmm5		; round to int32, mm5=data5L=(05 15)
+	cvtps2pi  mm4,xmm0		; round to int32, mm4=data3L=(03 13)
+	cvtps2pi  mm6,xmm4		; round to int32, mm6=data5H=(25 35)
+	cvtps2pi  mm7,xmm6		; round to int32, mm7=data3H=(23 33)
+	packssdw  mm5,mm6		; mm5=data5=(05 15 25 35)
+	packssdw  mm4,mm7		; mm4=data3=(03 13 23 33)
+
+	movq      mm6,[GOTOFF(ebx,PB_CENTERJSAMP)]	; mm6=[PB_CENTERJSAMP]
+
+	packsswb  mm2,mm3		; mm2=(02 12 22 32 04 14 24 34)
+	packsswb  mm4,mm5		; mm4=(03 13 23 33 05 15 25 35)
+
+	paddb     mm0,mm6
+	paddb     mm1,mm6
+	paddb     mm2,mm6
+	paddb     mm4,mm6
+
+	movq      mm7,mm0		; transpose coefficients(phase 1)
+	punpcklbw mm0,mm1		; mm0=(00 01 10 11 20 21 30 31)
+	punpckhbw mm7,mm1		; mm7=(06 07 16 17 26 27 36 37)
+	movq      mm3,mm2		; transpose coefficients(phase 1)
+	punpcklbw mm2,mm4		; mm2=(02 03 12 13 22 23 32 33)
+	punpckhbw mm3,mm4		; mm3=(04 05 14 15 24 25 34 35)
+
+	movq      mm5,mm0		; transpose coefficients(phase 2)
+	punpcklwd mm0,mm2		; mm0=(00 01 02 03 10 11 12 13)
+	punpckhwd mm5,mm2		; mm5=(20 21 22 23 30 31 32 33)
+	movq      mm6,mm3		; transpose coefficients(phase 2)
+	punpcklwd mm3,mm7		; mm3=(04 05 06 07 14 15 16 17)
+	punpckhwd mm6,mm7		; mm6=(24 25 26 27 34 35 36 37)
+
+	movq      mm1,mm0		; transpose coefficients(phase 3)
+	punpckldq mm0,mm3		; mm0=(00 01 02 03 04 05 06 07)
+	punpckhdq mm1,mm3		; mm1=(10 11 12 13 14 15 16 17)
+	movq      mm4,mm5		; transpose coefficients(phase 3)
+	punpckldq mm5,mm6		; mm5=(20 21 22 23 24 25 26 27)
+	punpckhdq mm4,mm6		; mm4=(30 31 32 33 34 35 36 37)
+
+	pushpic	ebx			; save GOT address
+
+	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
+	mov	ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
+	movq	MMWORD [edx+eax*SIZEOF_JSAMPLE], mm0
+	movq	MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm1
+	mov	edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
+	mov	ebx, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
+	movq	MMWORD [edx+eax*SIZEOF_JSAMPLE], mm5
+	movq	MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm4
+
+	poppic	ebx			; restore GOT address
+
+	add	esi, byte 4*SIZEOF_FAST_FLOAT	; wsptr
+	add	edi, byte 4*SIZEOF_JSAMPROW
+	dec	ecx				; ctr
+	jnz	near .rowloop
+
+	emms		; empty MMX state
+
+	pop	edi
+	pop	esi
+;	pop	edx		; need not be preserved
+;	pop	ecx		; need not be preserved
+	pop	ebx
+	mov	esp,ebp		; esp <- aligned ebp
+	pop	esp		; esp <- original ebp
+	pop	ebp
+	ret
+
diff --git a/common/jpeg/simd/jsimd.h b/common/jpeg/simd/jsimd.h
index 8e78eab..98bcebc 100644
--- a/common/jpeg/simd/jsimd.h
+++ b/common/jpeg/simd/jsimd.h
@@ -13,6 +13,7 @@
 #define JSIMD_NONE    0x00
 #define JSIMD_MMX     0x01
 #define JSIMD_3DNOW   0x02
+#define JSIMD_SSE     0x04
 
 /* Short forms of external names for systems with brain-damaged linkers. */
 
@@ -30,16 +31,22 @@
 #define jsimd_h2v1_merged_upsample_mmx        jSMUpH2V1M
 #define jsimd_convsamp_mmx                    jSConvM
 #define jsimd_convsamp_float_3dnow            jSConvF3D
+#define jsimd_convsamp_float_sse              jSConvFS
 #define jsimd_fdct_islow_mmx                  jSFDMIS
 #define jsimd_fdct_ifast_mmx                  jSFDMIF
 #define jsimd_fdct_float_3dnow                jSFD3DF
+#define jconst_fdct_float_sse                 jSCFDSF
+#define jsimd_fdct_float_sse                  jSFDSF
 #define jsimd_quantize_mmx                    jSQuantM
 #define jsimd_quantize_float_3dnow            jSQuantF3D
+#define jsimd_quantize_float_sse              jSQuantFS
 #define jsimd_idct_2x2_mmx                    jSIDM22
 #define jsimd_idct_4x4_mmx                    jSIDM44
 #define jsimd_idct_islow_mmx                  jSIDMIS
 #define jsimd_idct_ifast_mmx                  jSIDMIF
 #define jsimd_idct_float_3dnow                jSID3DF
+#define jconst_fdct_float_sse                 jSCIDSF
+#define jsimd_idct_float_sse                  jSIDSF
 #endif /* NEED_SHORT_EXTERNAL_NAMES */
 
 /* SIMD Ext: retrieve SIMD/CPU information */
@@ -96,12 +103,19 @@
                                              JDIMENSION start_col,
                                              FAST_FLOAT * workspace));
 
+EXTERN(void) jsimd_convsamp_float_sse JPP((JSAMPARRAY sample_data,
+                                           JDIMENSION start_col,
+                                           FAST_FLOAT * workspace));
+
 /* SIMD Forward DCT */
 EXTERN(void) jsimd_fdct_islow_mmx JPP((DCTELEM * data));
 EXTERN(void) jsimd_fdct_ifast_mmx JPP((DCTELEM * data));
 
 EXTERN(void) jsimd_fdct_float_3dnow JPP((FAST_FLOAT * data));
 
+extern const int jconst_fdct_float_sse[];
+EXTERN(void) jsimd_fdct_float_sse JPP((FAST_FLOAT * data));
+
 /* SIMD Quantization */
 EXTERN(void) jsimd_quantize_mmx JPP((JCOEFPTR coef_block,
                                      DCTELEM * divisors,
@@ -111,6 +125,10 @@
                                              FAST_FLOAT * divisors,
                                              FAST_FLOAT * workspace));
 
+EXTERN(void) jsimd_quantize_float_sse JPP((JCOEFPTR coef_block,
+                                           FAST_FLOAT * divisors,
+                                           FAST_FLOAT * workspace));
+
 /* SIMD Reduced Inverse DCT */
 EXTERN(void) jsimd_idct_2x2_mmx JPP((void * dct_table,
                                      JCOEFPTR coef_block,
@@ -136,3 +154,9 @@
                                          JSAMPARRAY output_buf,
                                          JDIMENSION output_col));
 
+extern const int jconst_idct_float_sse[];
+EXTERN(void) jsimd_idct_float_sse JPP((void * dct_table,
+                                       JCOEFPTR coef_block,
+                                       JSAMPARRAY output_buf,
+                                       JDIMENSION output_col));
+
diff --git a/common/jpeg/simd/jsimdcfg.inc.h b/common/jpeg/simd/jsimdcfg.inc.h
index e1f2bdd..2779565 100644
--- a/common/jpeg/simd/jsimdcfg.inc.h
+++ b/common/jpeg/simd/jsimdcfg.inc.h
@@ -98,6 +98,7 @@
 definev(JSIMD_NONE)
 definev(JSIMD_MMX)
 definev(JSIMD_3DNOW)
+definev(JSIMD_SSE)
 
 ; Short forms of external names for systems with brain-damaged linkers.
 ;
@@ -115,15 +116,21 @@
 definev(jsimd_h2v2_merged_upsample_mmx)
 definev(jsimd_convsamp_mmx)
 definev(jsimd_convsamp_float_3dnow)
+definev(jsimd_convsamp_float_sse)
 definev(jsimd_fdct_islow_mmx)
 definev(jsimd_fdct_ifast_mmx)
 definev(jsimd_fdct_float_3dnow)
+definev(jconst_fdct_float_sse)
+definev(jsimd_fdct_float_sse)
 definev(jsimd_quantize_mmx)
 definev(jsimd_quantize_float_3dnow)
+definev(jsimd_quantize_float_sse)
 definev(jsimd_idct_2x2_mmx)
 definev(jsimd_idct_4x4_mmx)
 definev(jsimd_idct_islow_mmx)
 definev(jsimd_idct_ifast_mmx)
 definev(jsimd_idct_float_3dnow)
+definev(jconst_idct_float_sse)
+definev(jsimd_idct_float_sse)
 #endif /* NEED_SHORT_EXTERNAL_NAMES */
 
diff --git a/common/jpeg/simd/jsimdcpu.asm b/common/jpeg/simd/jsimdcpu.asm
index b65699e..3561b08 100644
--- a/common/jpeg/simd/jsimdcpu.asm
+++ b/common/jpeg/simd/jsimdcpu.asm
@@ -66,6 +66,10 @@
 	jz	short .no_mmx
 	or	edi, byte JSIMD_MMX
 .no_mmx:
+	test	eax, 1<<25		; bit25:SSE
+	jz	short .no_sse
+	or	edi, byte JSIMD_SSE
+.no_sse:
 
 	; Check for 3DNow! instruction support
 	mov	eax, 0x80000000
diff --git a/common/jpeg/simd/jsimdext.inc b/common/jpeg/simd/jsimdext.inc
index 509e95d..5fcd7be 100644
--- a/common/jpeg/simd/jsimdext.inc
+++ b/common/jpeg/simd/jsimdext.inc
@@ -119,15 +119,23 @@
 %define SIZEOF_MMWORD           SIZEOF_QWORD    ; sizeof(MMWORD)
 %define MMWORD_BIT              QWORD_BIT       ; sizeof(MMWORD)*BYTE_BIT
 
+; NASM is buggy and doesn't properly handle operand sizes for SSE
+; instructions, so for now we have to define XMMWORD as blank.
+%define XMMWORD                                 ; int128 (SSE register)
+%define SIZEOF_XMMWORD          SIZEOF_OWORD    ; sizeof(XMMWORD)
+%define XMMWORD_BIT             OWORD_BIT       ; sizeof(XMMWORD)*BYTE_BIT
+
 %define SIZEOF_BYTE             1               ; sizeof(BYTE)
 %define SIZEOF_WORD             2               ; sizeof(WORD)
 %define SIZEOF_DWORD            4               ; sizeof(DWORD)
 %define SIZEOF_QWORD            8               ; sizeof(QWORD)
+%define SIZEOF_OWORD            16              ; sizeof(OWORD)
 
 %define BYTE_BIT                8               ; CHAR_BIT in C
 %define WORD_BIT                16              ; sizeof(WORD)*BYTE_BIT
 %define DWORD_BIT               32              ; sizeof(DWORD)*BYTE_BIT
 %define QWORD_BIT               64              ; sizeof(QWORD)*BYTE_BIT
+%define OWORD_BIT               128             ; sizeof(OWORD)*BYTE_BIT
 
 ; --------------------------------------------------------------------------
 ;  External Symbol Name